MMU-DPI: Enhancing Generalization in Drug–Protein Interaction Prediction through Multimodal Learning and a Label Mix Strategy

Abstract

Accurate prediction of drug–protein interactions (DPIs) is crucial for accelerating the drug discovery process. However, the scarcity of experimentally validated interactions can limit the learning of transferable interaction patterns, particularly for previously unseen drugs and proteins. To address this fundamental challenge, we propose the MMU-DPI framework. A key component of this framework is a Label Mix strategy tailored to multimodal DPI prediction, which performs interpolation only in the label space while keeping the input modalities unchanged. This strategy provides stochastic soft-target regularization and improves generalization performance under reduced-data and independent external Cold-both evaluation settings. To effectively process and utilize multimodal data, MMU-DPI adopts a multimodal dual-branch architecture. The first branch uses a Message Passing Neural Network (MPNN) to extract structured representations from drug molecular graphs. It also uses a Convolutional Neural Network (CNN) to capture key biological and functional features from amino acid sequences. The second branch constructs a heterogeneous interaction graph and uses a Graph Attention Network (GAT) to learn deep contextual relationships between drugs and proteins. A learnable global fusion weight combines complementary branch logits to generate the final prediction for each drug–protein pair. Experimental results on multiple benchmark data sets demonstrate that MMU-DPI outperforms several state-of-the-art DPI prediction methods. Case studies further support the ability of MMU-DPI to identify potential DPIs. These results indicate that MMU-DPI can serve as a useful computational tool for drug discovery.